---
id: cockpit-view
title: "First-Person Cockpit: the *_cop Canopy + the Authentic Eyepoint"
status: living
source_sections: "task #55 (2026-07-10); reference/decomp part_007.c/part_013.c/part_014.c/part_011.c; engine/MUNGA_L4/bgfload.cpp, L4VIDRND.cpp; game/reconstructed/btl4vid.cpp"
related_topics: [rendering, bgf-format, gauges-hud, subsystems, decomp-reference]
key_terms: [material-ramp, siteeyepoint]
open_questions: ["gyro tail: gait jolt/rumble [T3 gates], sway model, torso-pitch eye writer, damageForce fill", "exact authored frame colour: the unlit constant (0.13,0.12,0.15) is T3 tuned-to-footage (BT_COP_FRAME overrides; the ramp-index approach + BT_COP_RAMP_L are RETIRED)"]
---
# First-Person Cockpit: the *_cop Canopy + the Authentic Eyepoint

The pod's only view was the cockpit eyepoint looking out through a per-mech canopy shell.
Task #55 reconstructed both halves — the CANOPY rendering and the EYE camera — from the binary.
Verified in-game on all 8 mechs (Thor vs gameplay footage). [T2]

## The canopy shell (*_COP.BGF)

- One per mech (12: AVX BLX FIX JAX LOX MAX OWX RAX SNX STX THX VUX; model→skeleton table in
  BTL4.RES @~3232850: `madcat=mad`, `blkhawk=blh`, … — the INSIDE skeleton is the X-variant, so
  Madcat = `MAX_COP`). It is the torso segment's SkeletonType_A mesh; the inside view loads
  EXACTLY ONE segment mesh (the `_cop`) — 25 others hidden. [T2]
- Material: `<pfx>skin:blakskn_dz_{u,r,l}torso_mtl` — DIFFUSE/AMBIENT (0,0,0) + RAMP_REF
  `softer`, **NO texture** (BLXSKIN.BMF is 238 bytes total; byte-verified). The 56 nonzero-UV
  verts in BLX_COP sample nothing (planar-projection authoring leftover). [T1]
- Geogroups carry the `SV_SPECIAL " dz_*torso PUNCH "` token → `dpl_Punchize` (see
  [[bgf-format]] PUNCH). The board opcode is a constant — the frame/window discriminator is the
  GEOMETRY, not a key. [T1]
- **Every `_cop` is an OPEN STRUT LATTICE** (boundary-edge ratios, per-patch edge namespacing:
  38-59% across all 12). The struts ARE the frame; the openings ARE the windows. Beware the
  edge-counting trap: counting edges across patches (patch-local indices collide) undercounts
  boundaries catastrophically (BLX "7%" wrong → 59% right). [T1]
- **Rendering (FINAL treatment 2026-07-11; bgfload.cpp keyed by filename `meshIsCop`):**
  1. **The PUNCH STENCIL-CUT kit** (see §FINAL below + [[bgf-format]]): the TAG_PATCH scan
     hash-pairs the punch geogroup's chunks into mask / hull / twin roles; L4D3D executes the
     3-step stencil cut (D24S8). `BT_COP_PLATES=1` disables (diag).
  2. **Double-sided** like all no-normal geometry — the earlier single-siding + per-face inward
     winding existed solely to fight the solid-box look, which the un-cut PLATES caused, not the
     winding; both are REMOVED (bgfload.cpp:533 comment). `BT_COP_SINGLE=1` restores
     single-siding, `BT_COP_FLIP=1` flips its winding (diag only).
  3. **UNLIT constant frame colour** (the map-lighting bug fix, 2026-07-11): the frame batch is
     routed through the renderer's pure-emissive branch (black diffuse + emissive
     `(0.13,0.12,0.15)` [T3 tuned-to-footage]; `BT_COP_FRAME="r g b"` overrides). The previous
     lit path multiplied the colour by the MAP light — cavern-night `ambient=0.8` washed the
     frame gray (user-reported). `BT_COP_RAMP_L` + the flat-ramp-index approach are RETIRED.
- The shell SHOWS by default in the inside view (btl4vid.cpp ApplyViewSkeleton);
  `BT_HIDE_COCKPIT=1` hides it (diag).
- **Wrong turns, recorded so they stay dead:** (a) "windows are punch texels" — no texel exists;
  (b) "the punch patch IS the visible frame / drop-punch removes the cockpit" — an artifact of
  testing with the broken pre-inverse eye; (c) a view-direction face-cut heuristic — not
  data-derivable; (d) "drop the whole punch geogroup = the windows" and (e) "duplicate pairs =
  plates, hide both" — both superseded by the firmware-decoded STENCIL-CUT kit (§FINAL below),
  which is the authentic mechanism.

## The authentic eyepoint (decomp-verified [T1])

Pipeline: `eyeWorld = parentSegmentWorld · baseOffset · R(EyepointRotation)`; `VIEW =
inverse(eyeWorld)`. No LookAt anywhere.

- **Eye ctor = FUN_004579a8** (part_007.c:9274, = DPLEyeRenderable). Caller (part_014.c:5525-66,
  BTL4VID.CPP) gates on segment name `"siteeyepoint"`, passes:
  - offset matrix = the segment's **GetBaseOffset()** (segment+0x74, the LOCAL rest transform) —
    NOT GetSegmentToEntity, NOT an upright basis;
  - parent = the eyepoint segment's **parent draw component** (`dcs_array[GetParentIndex()]`) —
    liveness reaches the eye through this parent chain in the PORT TOO (task #58 FINAL): the
    twist path is `Torso::PushTwist → Joint::SetRotation → hinge rotationAmount →
    HingeRenderable::Execute` (multiplies the live hinge into the matrix stack), and the eye
    executes inside that scope (`jointtorso → jointeye → siteeyepoint`; VERIFIED: with the real
    jointtorso swept, the active cockpit `[eyefwd]` swings ~60° and the torso hinge renderable
    sits in the eye's parent chain). The CAMERA yaws with the twist, the canopy rides the same
    subtree, the crosshair stays screen-centered — see [[combat-damage]] Targeting.
    ⚠ TWO instrumentation traps produced false "eye is frozen" readings first: (1) the headless
    default camera is the CHASE eye (root-parented, same DPLEyeRenderable class — gate probes on
    the COCKPIT eye being live); (2) `BT_FORCE_TORSO` used to override the joint names with the
    BLH demo pair (`jointshakey2`/`jointtshadow`), silently redirecting a MadCat sweep into the
    SHADOW hinge — the hook now only fills joints the resource left unresolved. An explicit
    eye-side twist compose added on those false readings DOUBLE-ROTATED the view (2× vs the 1×
    canopy) and is retired;
  - the `EyepointRotation` entity attribute (EulerAngles; Mech member, init 0).
- **Combine order:** `FUN_0040b104(dst,A,B)` computes `dst = B × A` (translation from B) →
  the eye local = `baseOffset × R`, R applied FIRST in eye-local space. Recomputed only when the
  euler CHANGES (FUN_00457b48 @part_007.c:9360, epsilon gate).
- **Per-frame view = FUN_004c22c4** (part_013.c:11742): rebuilds eyeWorld from the view-DCS's
  live accumulated transform, then `FUN_0040b244` = TRUE AFFINE INVERSE → the view matrix, then a
  UNIFORM zoom scale. The forward/up axes fall out of the eye's own basis — the renderer never
  picks an axis. (Our old hand-rolled `D3DXMatrixLookAtRH` with forward=+Z/up=+Y row extraction
  was exactly why some mechs aimed into the canopy — see [[reconstruction-gotchas]].)
- Port implementation: btl4vid.cpp (eye build: offset_matrix + parent_DCS), L4VIDRND.cpp
  DPLEyeRenderable::Execute (order `Orient × R`, `D3DXMatrixInverse` on the live path — note the
  ctor also holds a copy of this code but its view write is DEAD; patch the Execute).

## Status per mech + the gyro (task #56 — gyro now LIVE)

- **Madcat (MAX_COP): verified good on pure defaults** — dark domed frame (top arch +
  cross-braces + A-pillars + dash), world through the openings. Un-regressed with the gyro live.
- **The gyro is RE-ENABLED and clean** (task #56): ctor field-map + integrators + writers
  reconstructed byte-exact from @004b3778/@004b2ec0/@004b30ec/@004b34ec, layout
  static_assert-locked (sizeof==0x3D0), joint-write dispatch moved to the MECH performance tail
  (`GyroFrameJointWrite`, matching the binary calls @0x4aaf74/83), gyro↔torso pitch link wired
  (`gyro+0x258 = &torso currentTwist`). Runtime: both joints resolve type 5, Performance
  installed (flags 0x100), `WriteMechJoint eyePos=(0,0,0) body=(0,0,0)` finite every frame — the
  old NaN is gone. Mechanism: WriteMechJoint drives ONE node (`jointeye`, which parents
  `siteeyepoint`): TRANSLATION = the eye position spring, ROTATION = the body tip spring.
  WriteEyeJoint is a MULTIPLICATIVE sway attenuator on `jointlocal`'s animated rotation
  (gated legAnimState!=0; must run after the animation pass). [T2 runtime-verified]
- **Steady-offset hypothesis DISPROVEN [T2]:** the .MDL spring targets are SYMMETRIC
  (pos=(0.1,0.15,0.1), neg=(−0.1,−0.15,−0.1) — same for madcat AND bhk1) → the eye equilibrium
  is (0,0,0) and the clamps limit gyro motion to ±0.1..0.15u. The gyro is a BOUNCE/sway
  mechanism, NOT a large steady offset. The authentic resting eye = raw baseOffset for every mech.
- **FINAL (2026-07-11, i860-firmware-decoded): PUNCH is a three-chunk STENCIL-CUT KIT.** The
  VREND.MNG board firmware's 'damageize' handler (@0xf040f6f8, disassembled with a custom i860
  tool) writes the dpl_Punchize token triple {dmg_set, undmg2enbl|texture, dmg_clear} onto the
  punch geogroup's first three geometry chunks IN FILE ORDER: chunk1 = the detailed
  window-aperture shapes = an invisible per-pixel MASK; chunk2 = the coarse HULL = the ONLY
  colour-drawn geometry, drawn where NOT masked; chunk3 = a byte-identical hull twin = the
  damage-reset record, never visible.  The visible canopy = hull-minus-apertures.  Port:
  bgfload identifies the within-patch twin pair (hull + skip) and tags the non-paired pmeshes as
  masks; L4D3D draws mask (stencil 1, no colour/z-write), hull (stencil-reject), mask (stencil 0)
  on a D24S8 device; the cut never touches z, so later-drawn entities show through apertures.
  BT_COP_PLATES=1 disables the kit (diag).  In-game, all 8 mechs on pure defaults [T2]:
  connected dark frames with clear viewports; thor (the footage mech) shows the big central
  viewport + surrounding frame exactly as filmed.  TWO earlier readings are superseded and were
  WRONG: "punch patches = windows, drop them" (left floating frame fragments) and "duplicate
  pairs = plates, hide both" (hid the HULL and drew the MASK = the inversion the user caught).
- **Hit-bounce: LIVE and verified [T2] (2026-07-11).** The damage→gyro fan-out **FUN_004b2980**
  was re-disassembled byte-exact from the raw image (capstone; scratchpad/dis_4b2980.py) and
  reconstructed as `Gyroscope::ApplyDamageResponse` (gyro.cpp): zero-damage + Collision(0) no-op;
  hit direction = `Damage::damageForce` or a RANDOM horizontal fallback when ~zero (our senders
  don't fill damageForce yet — the fallback is binary-legal; directional fidelity = fill it later);
  rotated by the yaw-only torso-twist frame; per-type scaling `amount / damageMultiplier[type] *
  damageResponse[type].{trans,pitchRoll,yaw,vibration}` (Explosive alone multiplies burstCount);
  clamp 1.3; four kicks — impulse(trans, dir), torque(pitchRoll, dir), impulse(vibration,
  **vibrationDirection@0x390 = (0,1,0)** — the "animationOffset/Scale/Phase scalars" were this
  unit vector all along), verticalImpulse(yaw, dir).  Call sites wired: the take-damage hub
  (mech.cpp TakeDamageMessageHandler, FIRST action, binary @0x4a0264), the crushable-icon crunch
  (mech4.cpp, torque 0.4 along Δv + up-impulse 0.2, @4aa81e/@4aa86c), the firing recoil
  (projweap.cpp FireWeapon: damage>3 → impulse (0,0.6,−1.5) × damage/16, @4bc136-4bc19c).
  Verified live: `[gyro-dmg]` fires per hit with correct scaling, eyePosition shows the damped
  oscillation (ramp to ~0.015u, Y oscillating, decay to 0), no NaN, kill chain un-regressed.
  **Quantitatively verified (2026-07-11 MP trace):** `BT_GYRO_TRACE=1` logs the integrator state
  every displaced frame (`[gtrace]`); a 25-hit autofire session (scratchpad/GYRO_TRACE.png)
  shows per-hit kick → damped oscillation → settle-to-zero-and-stay, peaks ±0.028u X/Z ±0.005u Y
  (well inside the ±0.1/0.15 clamps), body pitch/roll ±0.02 rad, zero NaN/drift — and the
  measured frequency ratio matches the authored constants: Y period 9.5 frames vs X/Z ~75 ≈
  √(springK.y/springK.x)=√60≈7.75× as dictated by springK=(−0.2,−12,−0.2). [T2]
- **Deferred gyro tail:** (a) the alternate-gait engage JOLT + 0.4s engaged-gait RUMBLE
  (@4aa158-4aa365 — full byte recipe in the wf_6880e605 synthesis; **flagged [T3] on the gate
  naming** and it mutates the gait state machine tasks #49/#50 stabilized — do NOT wire without
  re-verifying the gate semantics); (b) the mech+0x3F0 overspeed sway model (swayBias fed 0);
  (c) the per-frame `EyepointRotation.pitch = torso pitch` writer (FUN_004b66b4 @0x4aafab) + the
  mapper glance-look states (FUN_004afd10); (d) senders filling `Damage::damageForce` (beam:
  impact−muzzle; projectile: velocity) for directional bounce fidelity; (e) FireWeapon's heat add
  (pre-existing separate gap).  `gyroRumbleTimer@0x5c4` was re-typed from the mis-named
  `int clipLoadGuard` (binary uses it as a float rumble countdown).

## Offset cockpits — the placement-model verification [T1/T2]

Some mechs have laterally OFFSET cockpits, and the data + our port agree exactly (2026-07-11):
`[jointeye]` rest translation per SKL vs the `_COP` canopy X-center — **thor: tranx=+1.13, canopy
center +1.13 (exact)**; loki −0.25/−0.15; owens +0.19/+0.16; the other five ≈0/≈0.  Our in-game
measured eye (`[EYECHK]`) reproduces the Thor offset to the hundredth (lateral 1.13 from the mech
origin).  The eye chain (baseOffset + parent-segment DCS composition) is therefore CORRECT per-mech;
the canopy-render question was subsequently resolved by the stencil-cut kit — all 8 render
authentically (see the FINAL paragraph above).

## The playable roster is 8, not 12 [T1]

`BTL4.RES` contains exactly 8 mech model resources (avatar, bhk1/blkhawk, loki, madcat, owens,
sunder, thor, vulture — each with `<name>`, variants `<pfx>1/2`, and a `<pfx>dead`).  Strider /
Raptor / Firestarter / Blackjack have `_COP` meshes, SKLs and gauge configs but NO model resource
(cut/other-version leftovers); forcing them via `BT_FORCE_MODEL` feeds NULL to
`ResourceDescription::Lock` → AV in `CreatePlayerVehicle` (cdb-verified).  Valid `BT_FORCE_MODEL`
names: `avatar bhk1 loki madcat owens sunder thor vulture` (+ variants ava1/lok1/lok2/mad1/mad2/
own1/snd1/thr1/vul1/blkhawk).

## The authentic view field (FOV) — viewangle is HORIZONTAL [T1]

`BTDPL.INI [dpl_defaults] viewangle=60.0` is the **horizontal** field of view: the original Tesla
call (preserved in the DPLRenderer::SetView comment, L4VIDEO.cpp) is
`dpl_SetViewProjection(view, -1, -aspect, +1, +aspect, 1/tan(viewangle/2))` — horizontal
half-extent 1.0, vertical scaled by aspect.  Authentic frustum at 4:3 = **60° × 46.8°**.  The
port had fed the 60 to D3DX as the VERTICAL fov (75° × 60° at 4:3, worse on widescreen) — the
world read ~25% farther away ("camera too far back" reports) and the canopy frames were
massively over-revealed.  Fixed (`BTFovYFromHorizontal`, all 5 projection sites incl. sky +
resize): the horizontal field stays 60° at any window shape.  The aim/pick feed reads the live
projection (`gBTAimP22`) so targeting adapts automatically (combat re-verified).  Credit: a
user-community report ("frames were never this revealed") + the resolution/FOV hypothesis.

## Diagnostics / env gates

`BT_FORCE_MODEL=<name>` (force the player mech: madcat/owens/sunder/…; btl4mssn.cpp),
`BT_HIDE_COCKPIT`, `BT_COP_PLATES` (disable the stencil kit), `BT_COP_FRAME="r g b"` (frame
colour override), `BT_COP_SINGLE` / `BT_COP_FLIP` (single-siding diag), `BT_COP_DEBUG` (paint
the shell green), `BT_COP_DUMP` (per-batch punch/bbox), `BT_EYE_FWD=<f>` (eye position probe),
`BT_GYRO_LOG` / `BT_GYRO_TRACE` (gyro bring-up log / per-frame integrator trace, gyro.cpp),
`[EYECHK]` log (eye world pos + look dir + entity pos, first frames).
(RETIRED: `BT_COP_DOUBLE`, `BT_COP_RAMP_L` — removed with the stencil-cut/unlit-frame landing.)

## Key Relationships
- Uses: [[bgf-format]] (PUNCH, vertex tags) · [[rendering]] (ramp model, skeleton branch)
- Feeds: [[gauges-hud]] (the HUD overlays this view) · [[subsystems]] (gyro, task #56)
- Gotchas: [[reconstruction-gotchas]] §14 (LookAt axis guess), §15 (per-patch edge counting)